Generalized Mack Chain-Ladder

Model of Reserving
with Robust Estimation
by Przemyslaw Sloma

ABSTRACT
In this paper we consider the problem of stochastic claims reserv-
ing in the framework of development factor models (DFM). More
precisely, we provide the generalized Mack chain-ladder (GMCL)
model that expands the approaches of Mack (1993; 1994; 1999),
Saito (2009) and Murphy, Bardis, and Majidi (2012). Our general
flexible tool of reserving provides the solution to the one of the
major challenges of day-to-day actuarial practice, which is quantify-
ing the variability of reserves in the case where different methods of
selecting loss developments factors (LDFs) are applied. We develop
the theoretical background to estimate the conditional mean square
error of prediction (MSEP) of claims reserves that is consistent with
actuarial practice in selecting the LDFs.
Moreover, we present an example of GMCL’s application in
which we indicate how to bridge the estimation of parameters in
the chain-ladder framework with the robust estimation techniques.
Finally, we show how our approach can be used in validation of the
reserve risk evaluation in the Solvency 2 context.

KEYWORDS
Non-life insurance, stochastic claims reserving,
Mack chain ladder, robust estimation solvency 2

226 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 226 4/25/19 10:46 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

1. Introduction and motivation The adjustments to make data more homogeneous

are often justified for number of reasons: unstable
The provision for outstanding claims is one of the run-off triangles, outliers, inaccurate and incomplete
main components of technical provisions of insur- data, etc.,). Most actuaries use somewhat arbitrary
ance company’s liabilities. Measuring the deviation of rules of thumb in selecting the loss development
the true amount of reserves from its estimation is one factors (LDFs). Blumsohn and Laufer (2009) describe
of the major actuarial challenge. Senior managers, this topic in great detail. In this project, a group of
shareholders, rating agencies, and insurance regula- actuaries were asked to select LDFs for an incurred
tors all have an interest in knowing the magnitude of run-off triangle. The important number of ways of
these potential variations (reserve uncertainty) since LDFs selection was provided by the participants.
companies with large potential deviations need more The approaches proposed to evaluate the estimation
capital or reinsurance. of expected value of reserves varied widely and the
One of the most known method of reserving used additional information about the error of prediction of
in practice is the approach called chain-ladder. This this estimation could be helpful in decision making.
method belongs to the family of development factor That is why it is extremely important from a practi-
models (DFM). The first stochastic approach based cal point of view to have a method that provides the
on the chain-ladder technique was proposed by Mack estimation of conditional MSEP of ultimate claims
(1993; 1994). In these studies, Mack proposed the (or reserves) in the context of LDFs selection by
estimation of the mean square error of prediction actuaries. In the present study we provide such a tool
(MSEP) of claims reserves based upon all-year embedded in the theoretical framework to quantify
volume-weighted average of loss development factors the standard error of prediction of the claims reserves
(also called: link ratios, age-to-age factors, report-to- in the case where some factors have been excluded
report factors). The variance structure was supposed from the estimation of model parameters. However,
to be proportional to the development period’s initial we do not judge whether these ad hoc approaches
loss. This assumption is sufficient for the weighted of selecting factors are correct or wrong. We rather
average development factors to have optimal statisti- assume that the expert judgment taken by an actuary
cal properties (BLUE-best linear unbiased estimate). could always be justified by his specific knowledge
Some authors (see Murphy 1996, p. 188; Mack 1999, of considered business.
p. 15) pointed out that the estimation of chain-ladder Measuring the variability of the reserves in this
factors is connected with the estimation in the frame- context is poorly developed in the literature. That
work of linear model by weighted least squares is why in practice actuaries and reserving software
(WLS) regression approach. They also observed that developers often use the proxy methods based on
by modifying the original variance assumption from formula for MSEP derived in Mack (1993). The
Mack (1993) the corresponding estimators of chain- approximations mainly consist of replacing the main
ladder development factors keep their BLUE property parameters by theirs estimators computed by the other
(see Murphy 1996; Barnett and Zehnwirth 2000; approach without changing the main formula. This
Saito 2009). It is worth underlying that the modi- procedure is incorrect because in the chain-ladder
fication of variance assumption leads to different framework the formula for MSEP depends among
point estimators for development factors (arithmetic others on the standard error of chain-ladder factors
average, slope of regression etc.; see Remark 3.1 for and it is not accurate to simply plug-in the new esti-
more details). mators in the old formula. The other proxy method
One of the major challenges in everyday actuarial often used in practice consists on applying the coeffi-
practice is selecting the loss development factors. cients of variation of ultimate loss from Mack (1993)

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 227

15080-07_Sloma-2ndPgs.indd 227 4/25/19 10:46 AM

 Variance Advancing the Science of Risk

(ratio of square root of MSEP of ultimate loss over type stochastic reserving methods to determine the
ultimate loss) in order to derive the MSEP estimators economic capital corresponding to the reserve risk.
of a new approach. It turns out that in general these The bootstrap method allows estimation of a whole
approximations are highly inappropriate (see exam- claims reserves distribution via resampling tech-
ple in Section 6.3). niques and Monte Carlo simulations. It seems to
We think that, in some simple cases (no curve be crucial for non-life insurance companies to be
fitting for LDFs, for example) the approximations able to validate the results given by the industrial
mentioned above are the consequences of bad under- software where we do not have access to the code
standing of the main formula for estimation of MSEP and when the number of shortcuts may be applied.
in Mack (1993). Moreover, the approximations used Our approach can be used to validate the estimation
by actuaries and actuarial software developers could of the first two moments of the loss distribution in
be avoided by using the more appropriate existing the case where selection of development factors was
models. One such model was proposed by Mack employed and the different weights in estimation of
(1999). To our knowledge, this was the first study that chain-ladder factors and volatility parameters were
showed how to measure the uncertainty of reserves used (see Section 6.3 for more details).
in the situation when an actuary selects the LDFs. Second, our general Mack chain-ladder (GMCL)
In our opinion, the important results obtained in this model can be used to construct the proxy solutions
paper are not always used in practice because, instead to overcome the limits of Mack’s (1999) approach,
of explicit formula, the recursive equation is given i.e., the use of the same weights for parameters esti-
there for estimation of MSEP of ultimate loss. mation (see discussion in Section 3.6). This means
Mack (1999) is an important paper that allows that, for the methods where we eliminate the consid-
the fully understanding of the MSEP formula and erable number of observations, we reduce as well the
avoid the inappropriate approximation when it is not data for variability of the reserves. This mechanically
necessary. We summarize the details of this method impacts the estimation of MSEP of loss liabilities,
in Section 3. One of the major limitations of this which in such cases is generally underestimated. We
method is the underestimation the MSEP of ultimate propose then the possible solution to overcome this
loss in the case where the number of excluding kind of difficulty (see Section 6.1 for more details).
data is important. We discuss this topic in detail in Finally, the third and really important application
Section 3.5. One possible solution to overcome this from a practical point of view consists of bridg-
difficulty is to extend the existing approach proposed ing the point estimation of chain-ladder parameters
by Mack (1999). with the theory of robust statistics. As mentioned
Therefore, we propose a general approach for above, the point estimators of chain-ladder factors
stochastic claims reserving in the framework of chain- can be obtained in the linear regression framework
ladder model, extending the model proposed by Mack by applying the weighted least squares procedure. It
(1999) and Murphy, Bardis, and Majidi (2012). This is well known that the OLS estimators are fragile to
extension is three fold. First, our general tool has a the outliers. That is why we propose using the robust
educational role and makes it possible to validate the techniques of estimation such as: M-estimators,
results from other approaches. More precisely, our Lp-estimators, etc (see Section 6.2).
general formula for estimation of MSEP of outstand- The reminder of this paper is organized as follows.
ing loss liabilities can be used to fully understand the In Section 2 we present our notations and definitions.
Mack (1993), Mack (1994), and Mack (1999) model. We review in Section 3 the MCL and its main limi-
Furthermore, under new solvency requirements of tations. In Section 4, we present the (GMCL) and
Solvency II, insurance companies use the bootstrap- the main results are derived in Section 5. Finally,

228 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 228 4/25/19 11:24 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

Section 6 introduces the numerical applications of Ri = Ci,I − Ci,I −i +1 , (2.1)

GMCL. All proofs are provided in the Appendix.
The related topics such as: tail factor, curve fitting, and the total outstanding loss liabilities for all acci-
diagnostics and validation of the main model hypoth- dent years,
esis, are out of scope of this paper and will be treated
I
elsewhere. R = ∑ Ri . (2.2)
i =1

2. Notations and definitions We use the term claims reserves to describe the pre-
diction of the outstanding loss liabilities. Hence, let
2.1. Run-off triangle
R̂i and R denote the claims reserves for accident year i,
Let Ci,j denote the random variables (cumulative R̂i = Ĉi,I – Ci,I–i+1, i ∈ {1, . . . , I}, and the total claims
payments, inccured, reported claims numbers, etc.) reserves for aggregated accident years, R̂ = ∑ i=1 I
R̂i,
for accident year i ∈ {1, . . . , I} until development year respectively, where Ĉi,I is a predictor for Ci,I.
j ∈ {1, . . . , J}, where the accident year is referred
to as the year in which an event triggering insurance 2.3. (Conditional) mean square error
claims occurs. We assume that Ci,j are random vari- of prediction (MSEP)
ables observable for calendar years i + j ≤ I + 1 and As already stated above, finding suitable pre­
non-observable (to be predicted) for calendar years diction of ultimate loss is rather the beginning of
i + j > I + 1. The observable Ci,j are represented by the process of reserving, and insurers need to assess
the so-called run-off trapezoids (I > J) or run-off tri- the variability of these amounts. We are interested then
angles (I = J ). Table 1 gives an example of a typical in the quantification of the prediction uncertainty of
run-off triangle. In order to simplify our notation, the ultimate loss, i.e., Ĉi,I and ∑ i=1
I
Ĉi,I, (or equivalently
we assume that I = J (run-off triangle). However, all of claims reserves, i.e., R̂ i and R̂ = ∑ i=1 I
R̂i). For that,
the results we present here can be easily extended to we have to choose an appropriate risk measure which
the case when the last accident year for which data determines a conception of measuring the “distance”
is available is greater than the last development year, between the prediction and the actual outcomes.
i.e., I > J (run-off trapezoid). In this paper, following the actuarial literature, we
quantify the prediction uncertainty using the most
2.2. Outstanding reserves popular such measure, the so-called mean-square
Let Ri et R denote the outstanding claims liabilities error of prediction (MSEP).
for accident year i ∈ {1, . . . , I},
( )
2
msepCˆiI DI (CiI ) = E  Cˆ iI − CiI DI  , (2.3)
 
Table 1. Run-off triangle (I = J )
2
Development Year j  I   I ˆ I
 
Accident
msep ˆ iI DI  ∑ C
 i =1 iI  = E   ∑ C − ∑ CiI  DI  ,
∑ i =1  i =1 
I iI
Year i 1 2 3 4 j ... J C
i =1 
1
(2.4)
2 Ci,j
(observations)
3 where
I–j
I–2 Ci,j DI = {Ci, j : i + j ≤ I + 1} , (2.5)
(to be predicted)
I–1
I denote the claims data available at time t = I.

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 229

15080-07_Sloma-2ndPgs.indd 229 4/25/19 10:46 AM

 Variance Advancing the Science of Risk

3. Mack chain-ladder (MCL) model Var ( Fi,k Ci,1 , . . . , Ci,k ) =

σ 2k
,
wi,k Ciα,k
A major everyday challenge of actuarial work is
selecting loss development factors for number of with wi,k ∈[ 0, 1]. (3.2)
reasons (outliers in triangle, inaccurate data, incom-
pleteness, etc). Most actuaries use somewhat arbi- The parameters sk are referred here as variance
trary rules of thumb in selecting the loss ratios. In parameters (LDF).
Blumsohn and Laufer (2009), a group of actuaries 3. The accident years (Ci,1, . . . , Ci,I)1≤i≤I are indepen­dent.
were asked to select age-to age factors for a 12-years
triangle of umbrella business. The important number 3.2. Estimation of parameters
of ways of selecting loss ratios was provided by the in the MCL model
participants. It is important, then, from practical point
• Given the information DI and for 1 ≤ k ≤ I – 1, the
of view to have a method that provides the estima-
factors fk are estimated by
tion of conditional MSEP of ultimate claims in the
context of factor selection of actuaries. I −k
To the best of our knowledge, the paper by Mack ∑ wi,kCiα,k Fi,k
(1999) is one of the first studies dealing with factors fˆk = i =1
I −k , α ∈{0, 1, 2} . (3.3)
selection and variability of reserves estimation in the ∑ γ i,k
i =1
framework of the chain-ladder method. This paper is
an extension of Mack (1993). • Given the information DI and for 1 ≤ k ≤ I – 2, the
In the remaining part of this section we recall variance parameters s k2 are estimated by
the assumptions of the Mack chain-ladder (MCL)
model from Mack (1999). Afterwards, we present 1 I −k
( )
2
the numerical example illustrating the limits of this σˆ 2k = ∑ wi,k Ciα,k Fi,k − fˆk , α ∈{0, 1, 2} ,
I k − 1 i =1
approach. Finally, we indicate the possible expansion (3.4)
of MCL method and its potential applications.
where Ik represents the number of weights wi,k different
3.1. Model assumptions of MCL method from 0, namely, Ik := card{i: wi,k ≠ 0}.
Let define the individual development factors, for Formula (3.4) does not yield an estimator for ŝ 2I–1
1 ≤ i ≤ I – 1 and 1 ≤ k ≤ I – 1, because it is not possible to estimate this param-
eter from the single observation CI,I /CI,I–1. Follow-
Fi,k = Ci,k +1 Ci,k . (3.1) ing Mack (1993; 1994; 1999), if fI–1 = 1 and if the
claims development is believed to be finished after
Following Mack (1999), we assume [(MCL.1)] I – 1 years we can put ŝ 2I–1 = 0. If not, the simple
1. There exist constants fk > 0 such that formula of extrapolation can be applied by requir-
ing ŝI–3/ŝI–2 = ŝI–2 /ŝI–1. This leads to the following
E ( Fi,k Ci,1 , . . . , Ci,k ) = fk . definition

The parameters fk are often called loss devel- σˆ 2I −1 := min ( σˆ 4I − 2 σˆ 2I −3 , min ( σˆ 2I −3 , σˆ 2I − 2 )) . (3.5)
opment factors (LDF), link ratios or age-to-age
factors. Remark 3.1. The parameter α determines the
2. There exist constants s k2 > 0 such that for all different ways of estimation of fk . For the sake of
1 ≤ i ≤ I and 1 ≤ k ≤ I – 1 we have simplicity, let us assume that wi,j = 1 for all i, j.

230 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 230 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

We present below the possible choices of α and their This implies, together with the fact that fˆk are
interpretation. uncorrelated, that Ĉi,I is unbiased estimator of
E(Ci,I DI).
1. If α = 1 we get the classical chain ladder estimate
v. The expected values of the estimator
of fk
I −1
I −k I −k Cˆ i,I = Ci,I +1−i i ∏ fˆk ,
∑ Ci,k Fi,k ∑ Ci,k +1 k = I +1− i

fˆk = i =1
I −k = i =1
I −k , for 1 ≤ k ≤ I − 1.
for the ultimate claims amount and of the true ulti-
∑ Ci,k ∑ Ci,k
i =1 i =1 mate claims amount Ci,I are equal, i.e., E(Ĉi,I) =
E(Ci,I), 2 ≤ i ≤ I.
2. If α = 0 we get the model for which the estimators
The proof is postponed to Appendix A.3.
of the age-to-age factors fk are the straightforward
average of the observed individual development
3.4. Estimators of conditional MSEP
factors Fi,j defined via (3.1), i.e.,
in MCL model
1 I −k 3.4.1. Single accident years
fˆk = ∑ Fi,k , for 1 ≤ k ≤ I − 1.
I − k i =1 Under assumptions of the MCL model we have the
following estimator for the conditional estimation
3. If α = 2 we get the model for which the estima- error of a single accident year i ∈ {2, . . . , I}:
tors of the age-to-age factors fk are the results of
 ˆ ( Ci , I )
msep
an ordinary regression of {Ci,k+1}i∈{1,...,I–k–1} against Ci,I DI

{Ci,k}i∈{1,...,I–k} with intercept 0, i.e.,

I −1
σˆ 2k  1
( ) 1 ,
2
= Cˆ i ,I i ∑ ˆ2 ˆ ˆα + I −k (3.6)
I −k I −k k = I − i +1 f k  w C 
∑ Ci2,k Fi,k ∑ Ci,kCi,k +1  i,k i ,k ∑ wj,kC αj,k 
j =1
fˆk = i =1
I −k = i =1
I − k −1 , for 0 ≤ k ≤ I − 1.
∑C 2
i, k ∑ C 2
i, k where, for i + k > I + 1, we define ŵi,k : = 1 and fˆj and
i =1 i=0
ŝ 2j are given in (3.3) and (3.4)–(3.5) respectively.

3.3. Properties of estimators

from MCL model 3.4.2. Aggregated accident years
 I  I 
i, I  = ∑ msepCˆ i,I DI (Ci, I )
Proposition 3.1  I
msep  ∑ C
∑ Cˆ DI
i =1 i, I  i=2
i =1
i. The estimators fˆk given in (3.3) are unbiased and
( )
2
 I ˆ  I −1 σˆ k fˆk
2
uncorrelated. I
ˆ
+ ∑ Ci,I  ∑ C j,I  ∑ 2 I − k , (3.7)
ii. The estimators fˆk of fk have the minimal variance  j =i +1  k = I −i +1
i=2
∑ l,k l,k
w C α

among all unbiased estimators of fk which are the l =1

weighted average of the observed development

factors Fi,k. where fˆj and ŝ 2j are given in (3.3) and (3.4)–(3.5)
iii. The estimator ŝ 2k , given in (3.4) is the unbiased respectively.
estimator of the parameter s 2k .
iv. 
Under the model assumptions (MCL.1) and 3.5. Numerical application
of the MCL method
(MCL.3) we have
As mentioned above, the factors selection methods
E (Ci,I DI ) = Ci,I +1−i fi,I +1−i i . . . i fI −1 . are an integral part of everyday actuarial practice.

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 231

15080-07_Sloma-2ndPgs.indd 231 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

Here we choose from Blumsohn and Laufer (2009) (5) Median: fˆk are computed as an arithmetic aver-
several such methods where estimates are com- age of individual link ratios Fi,j in the way to
puted as different averages using varying weights obtain the sample median. More precisely, we
and varying number of accident years: all/3/5-years put wi,j = 1 or wi,j = 0 in the way that the estima-
weighted average and all excluding higher and lower tors of the age-to-age factors fk are given by
(AEHL) factor average. We consider as well other
popular methods in actuarial practice based on sam- fˆk = median {Fi,k : i ∈{1, . . . , I − k }} .
ple median.
More precisely, for RAA run-off triangle (see The median denotes the sample median that
Appendix B, Section B.8), we apply the MCL model for the sample X1, . . . , Xn is computed by
from Mack (1999) with the following parameters.
median { X i : i ∈{1, . . . , n}}
For all five methods described below we choose
α = 0 ( fˆk arithmetic averages of Fi,k) and we compute  X  n +1 if n is odd
the estimators fˆk and ŝ 2k according to formula (3.3)   2 

and (3.4)–(3.5), respectively. This allows us to com- :=  X + X .
 n n 
  2   +1
2
pare the results with sample Median method which  otherwise
2
is rather consistent with straightforward average of
development factors (see method number (5) below) where X(k) denotes the k – th order statistics of the
(1) ALL AV: fˆk are computed as arithmetic average sample X1, . . . ,Xn.
of all individual link ratios Fi,j. More precisely,
Remark 3.2. In the case where there is only
we define the weights in the following way:
one observation in estimation of parameter sk (odd
wi,j = 1 for all i, j.
number of data in sample median computation) we
(2) AEHL: fˆk are computed as arithmetic average of choose the additional Fi,k factor in order to have two
all individual link ratios, excluding the highest observations and be able to apply the formula (3.4).
and the lowest values of Fi,j. More precisely,
we define the weights in the following way: In Table 2, we present the estimation of total amount
for fixed j, wi,j = 0 for i such that Fi,j = F(I–j),j and of claims reserves R̂ as well as the value of estimators
Fi,j = F(1),j, where F(k),j for k = 1, . . . , I – j denotes of aggregated MSEP(R̂). Recall that R̂ : = ∑i=1 I
R̂ i,
the order statistics of Fi,j. For remaining indices i, where R̂i := Ĉi,I – Ci,I–i+1. We observe that to obtain R̂ it
for fixed j, we take wi,j = 1 is enough to have the estimators Ci,I of ultimate claims
(3) 5 Years AV: fˆk are computed as an arithmetic Ĉi,I for all accident year i. In consequence MSEP(R̂) =
I
average of individual link ratios Fi,j from five MSEP(∑ i=1 Ĉi,I) and we use the formula (3.7) to esti-
latest accidents years. More precisely, we define mate this quantity. We compute as well the coefficient
the weights in the following way: wi,j = 1 for of variation of R̂, given by CV(R̂) = R̂/MSEP(R̂)1/2.
i = I – j, . . . , I – j – 4. For remaining indices i, The last two lines of Table 2 indicate the relative
for fixed j, we take wi,j = 0 proportion of R̂ and MSEP(R̂)1/2, for each of five
methods considered, in comparison to the ALL AV
(4) 3 Years AV: fˆk are computed as an arithmetic
method which is the reference method in our example.
average of individual link ratios Fi,j from three
latest accidents years. More precisely, we define
the weights in the following way: wi,j = 1 for
3.6. Limits of MCL method
i = I – j, . . . , I – j – 2. For remaining indices i, As can be seen in Table 2, the four last methods
for fixed j, we take wi,j = 0 (columns (2)–(5)) reduce significantly the estimation

232 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 232 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

Table 2. Estimation of total amount of outstanding loss liabilities (R̂), value of estimator
of aggregated MSEP(R̂)1/2 and coefficient of variation CV(R̂), for five methods
alpha=0
ALL AV AEHL 5 Years AV 3 Years AV Median
Item/method (1) (2) (3) (4) (5)
R̂ 93 643 65 868 75 886 68 645 54 059
MSEP(R̂) 1/2
92 549 21 015 27 486 29 493 14 786
CV(R̂) 99% 32% 36% 43% 27%
(1)/(1) (2)/(1) (3)/(1) (4)/(1) (5)/(1)
R̂ (%) 100% 70% 81% 73% 58%
MSEP(R̂)1/2 (%) 100% 23% 30% 32% 16%

of MSEP(R̂)1/2 comparing to the first method ALL AV. for the expectation of fˆk and Var( fˆkBk) are not correct.
For the methods (3), (4) and (5), this is mainly due Regarding the sample median method, the derivation
to the elimination of relatively significant number of of Var( fˆkBk) requires the computation of the moments
development factors from estimation especially for of order statistics (see Jeng 2010) and those are
the first development years which correspond to the strongly related to the distribution of Fi,j. To over-
columns of the run-off triangle. This phenomena is come these difficulties we propose two solutions:
especially seen in the case of sample median method the simple Proxy method (see Section 6.1) and the
in which, for each development factor we keep at more complex one based on a robust estimation (see
most two of link ratios Fi,j in estimation of fk. From Section 6.2). The first approach is programmed to
statistical point of view, this is clearly not enough avoid artificial volatility increase and it is based on
to perform the robust estimation. As a consequence, all link ratios in estimation of volatility parameters sk
this kind of methods reduce unnaturally the variabil- (scale parameters in linear regression). The second
ity of reserves. This could be dangerous for exam- method consists on developing an approach that allows
ple in terms of evaluation of the economical capital using any robust estimators of fk (location) and sk
for reserve risk required by the new Solvency II (scale) parameters.
regime.
Beyond the limits stated above, there are some
incoherences with application of weights wi,k for 4. General Mack chain-ladder
the AEHL and sample median methods. Indeed, the model
weights wi,k should be Ci,k measurable random vari-
ables in order to be able to derive the main results of 4.1. Model assumptions
MCL approach (see, for example, Proposition A.2). Before stating the main assumptions of our general
Although for the method 5 Year AV and 3 Years AV approach, let us assume that functions gd,j:[0, ∞) →
we can fix the weights without knowing the infor­ [0, ∞) are Borel measurable. Let di,j be the non-
mation DI (knowing all observation in the run-off
negative random variables defined by, di,j: = gd,j(Ci,j).
triangle, see (5)), this is not a case for the AEHL and
Our model is formalized by the following
sample median methods. The reason is that we need
assumptions:
to know the observation Fi,k in order to specify the
corresponding weights for those two methods. That (GMCL.1) There exist constants fk > 0 such that
is why the weights wi,k are not Ci,k measurable but
rather DI-measurable. This means that the formula E ( Fi,k Ci,1 , . . . , Ci,k ) = fk .

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 233

15080-07_Sloma-2ndPgs.indd 233 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

(GMCL.2) There exist constants s 2k > 0 such that where Ik represents the number of weights di,k
for all 1 ≤ i ≤ I and 1 ≤ k ≤ I – 1 we have different from 0, namely, Ik := card{i:di,k ≠ 0}.
In the analogue way to (3.4) we define
 σk
2

 if δ i,k ≠ 0 a.s.,
Var ( Fi,k Ci,1 , . . . , Ci,k ) =  δ i,k σˆ 2I −1 = min ( σˆ 4I − 2 σˆ 2I −3 , min ( σˆ 2I −3 , σˆ 2I − 2 )) . (4.5)

∞ if δ i,k = 0 a.s.,
Proposition 4.1.
(4.1)
(i) The estimators fˆk given in (4.2) are unbiased and
uncorrelated.
where a.s. means almost surely.
(ii) For k = 1, . . . , I – 1, if di,k = γi,k for all i, then
(GMCL.3) The accident years (Ci,1, . . . , Ci,J)1≤i≤I
the estimators fˆk of fk have the minimal variance
are independent
among all unbiased estimators of fk which are
We observe that from the above assumptions the the weighted average of the observed develop-
main difference between MCL and GMCL lies in ment factors Fi,k.
the variance assumption. This modification allows For k = 1, . . . , I – 1, if di,k ≠ γi,k, for some i,
us to introduce different weights in estimation of the then the relative efficiency of s.e. ( fˆkg≠dBk) with
parameters fk and sk. respect to s.e.( fˆkg=dBk), i.e., the ratio

4.2. Model estimators I −k

γ2
(
s.e. fˆ γ =δ
Bk )Var fˆ (
Bk γ ≠δ
)
12 ∑ δ j,k i 1{δ
j, k ≠ 0 }
Suppose that functions gg,j:[0, ∞) → [0, ∞) are Borel k
:=
k
=
j =1 j, k

measurable. Let γi,j be the non-negative random vari- (

s.e. fˆ kγ ≠δ Bk ) (
Var fˆ kγ =δ Bk )
12 I −k
∑ γ j,k i 1{δ j, k ≠ 0 }
j =1
ables defined by, γi,j := gg,j(Ci,j).
• Given the information DI, the factors fk are esti- (iii) For k = 1, . . . , I – 1, if di,j = γi,j for all i, then
mated by the estimator ŝ k2, given in (4.4) is the unbiased
estimator of the parameter s k2.
I −k
∑ γ i,k Fi,k For k = 1, . . . , I – 1, if di,k ≠ γi,k, for some i,
fˆk = i =1
I −k , for 1 ≤ k ≤ I − 1 (4.2) then the bias of the estimator ŝ k2 is given by the
∑ γ i,k following formula
i =1

 I − k  I − k γ 2j,k  
It becomes obvious from assumption (GMCL.2)  ∑ δ i,k  ∑ i 1{δ }
j, k ≠ 0  
σk2
 j =1 δ j,k 
that in order to compute correctly the variance E [ σˆ 2k − σ 2k ] = E .
i =1
− 1
Ik − 1  
2
of fˆk (see Proposition A.2 in Appendix) we have  I −k 
  ∑ γ j,k  
to assume that  j =1 

{if δ i, j = 0 then γ i, j = 0} . (4.3) (iv) Under the model assumptions (GMCL.1) and
(GMCL.3) we have
• Given the information DI, the variance parameters
s 2k are estimated by E (Ci,I DI ) = Ci,I +1−i fi,I +1−i i . . . i fI −1 .

1 I −k This implies, together with the fact that fˆk are

( )
2
σˆ 2k = ∑ δ i,k Fi,k − fˆk , for 1 ≤ k ≤ I − 2,
I k − 1 i =1 uncorrelated, that Ĉi,I is unbiased estimator of
(4.4) E(Ci,IDI).

234 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 234 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

(v) The expected values of the estimator   I  I 

msep ∑ Cˆ DI
i =1 iI
 ∑ C i, I  = ∑ msepCˆ iI DI (CiI )
 i=2
i =1
I −1
Cˆ i,I = Ci,I +1−i i ∏ fˆk ,
( )
2
 I ˆ  J −1 σˆ k fˆk
2
k = I +1− i I
ˆ
+ ∑ Ci,I  ∑ C j,I  ∑ 2 I − k
 j =i +1  k = I −i +1  
2

 ∑ γ l,k 
i=2
for the ultimate claims amount and of the true
ultimate claims amount Ci,I are equal, i.e., l =1

E(Ĉi,I) = E(Ci,I), 2 ≤ i ≤ I. I −k
( γ l,k )2 i 1
i ∑ δ l,k {δ j,k ≠ 0} , (5.4)
The proof of this Proposition is postponed to the l =1

Appendix A.4.
where fˆj and ŝ 2j are defined in (4.2) and (4.4)–(4.5),
Remark 4.1.
respectively.
• If we set γi,j = di,j = wi,jC αi,j, for α ∈ {0,1,2}, in (4.2)
and (4.4) we get the assumptions of MCL model
from Mack (1999) (see also Mack (1993), Mack 6. Applications of GMCL model
(1994) and Saito 2009).
• If we put γi,j = di,j = wi,jC αi,jj , for αj ∈ , in (4.2) In our numerical example in Section 3.5 we have
and (4.4) we get the stochastic chain-ladder model seen that the assumption about the same weights in
from Murphy, Bardis and Majidi (2012). estimation of parameters sk and fk yields for some
methods to an artificial reduction of variability of
reserves amounts (refer to Table 2). To overcome
5. Main results this difficulty, we introduced the different weights
5.1. Single accident years γi,j and di,j in computation of fˆk and ŝk, respectively.
In the following application we indicate how one can
Result 5.1 (Conditional MSEP estimator for a possibly estimate the weights γi,j and di,j and we point
single accident year). out some other interesting applications.
 ˆ (Ci,I ) = Cˆ i,I ( ) i ( Γˆ )
2
msep i, I + ∆ˆ i,I , (5.1)
6.1. Method proxy for factors selection
Ci,I DI

where In this section, we examine our general framework

gg,j(Ci,j) := w γi,jC αi,j and gd,jCi,j := w di,j C βi,j which means
( fˆ )
2
J −1 σˆ 2k that γi,j := w γi,jC αi,j and di,j := w i,jd C βi,j. In this so called
Γˆ i,I = ∑
k
, (5.2)
k = I − i +1 δˆ i,k proxy method we impose using all link ratios Fi,j in
estimation of parameters sk (w i,jd = 1, for all i, j). For
( fˆ )
2

∆ˆ i,I =
J −1 σˆ 2k I −k
( γ j,k )2 i 1 all five methods presented, we take α = β = 0. We turn
∑
k
2
i ∑ δ j,k {δ j,k ≠ 0} , (5.3)
back to our numerical example from Section 3.5 and
k = I − i +1  I −k  j =1
 ∑ γ j,k  we evaluate the same estimators for the already-
j =1
presented five methods with the only difference in
and fˆj and ŝ 2j are given in (4.2) and (4.4)–(4.5), weights of sk estimation. More precisely:
respectively.
ALL AV: The ŝ 2k are estimated with w i,jd = 1 for
(1) 
all i, j. Parameters fˆk are computed as a arithmetic
5.2. Aggregation over prior accident year average of all individual link ratios Fi,j. More
Result 5.2 (Conditional MSEP estimator for precisely, we define the weights in the following
aggregated years). way: w γi,j = 1 for all i, j.

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 235

15080-07_Sloma-2ndPgs.indd 235 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

AEHL: The ŝ 2k are estimated with w i,jd = 1 for all

(2)  where median denotes the sample median which,
i, j. Parameters fˆk are computed as a arithmetic for the sample X1, . . . , Xn, is computed by
average of all individual link ratios excluding the
highest and the lowest values of Fi,j. More pre- median { X i : i ∈{1, . . . , n}}
cisely, we define the weights in the following way: if n is odd
 X  n +1
for fixed j, w i,jγ = 0 for i such that Fi,j = F(I–j),j and   2 
Fi,j = F(1),j, where F(k),j for k = 1, . . . , I – j denotes 
:=  X + X
 n n 
the order statistics of Fi,j. For remaining indices i,   2   +1
2
 otherwise
for fixed j, we take wi,jγ = 1. 2
(3) 5 Years AV: The ŝ 2k are estimated with w i,jd = 1
for all i, j. Parameters fˆk are computed as an and X(k) denotes the kth order statistics of the
arithmetic average of individual link ratios Fi,j sample X1, . . . ,Xn.
from five latest accidents years. More precisely, In Table 3 we present the estimation of R̂ and
we define the weights in the following way: MSEP(R̂) using the five methods described above.
w i,jγ = 1 for i = I – j, . . . , I – j – 4. For remaining In terms of MSEP we see that, in general, we have
indices i, for fixed j, we take w i,jγ = 0 the values greater than our reference method ALL AV
(4) 3 Years AV: The ŝ 2k are estimated with w i,jd = 1 from column (1), which stays unchanged compared
for all i, j. Parameters fˆk are computed as an arith- to Table 2. This is not surprising because by selecting
metic average of individual link ratios Fi,j from of the development factors we deacresed the estimated
three latest accidents years. More precisely, we values of fk and by using all observations Fi,j in ŝk
define the weights in the following way: w i,jγ = 1 computation we mechanically increased the disper-
for i = I – j, . . . , I – j – 2. For remaining indices i, sion around the values of fˆk. In view of our results
for fixed j, we take w i,jγ = 0 from Tables 2 and 3, the proxy method overestimates
(5) Median: The ŝ 2k are estimated with w i,jd = 1 for all in general the real MSEP, and can then be treated
i, j. Parameters fˆk are computed as an arithmetic as its upper bound. However, it can be useful as a
average of individual link ratios Fi,j in the way tool to perform the sensitivity analysis for testing the
to obtain the sample median. More precisely, we impact on the reserve volatility of excluding the
put w i,jγ = 1 or w i,jγ = 0 in the way that the estima- specific set of link ratios. Finally, it can be seen as
tors of the age-to-age factors fk are given by a measure of relative prudence of other approach of
measuring the variability of reserves by means of
fˆk = median {Fi,k: i ∈{1, . . . , I − k }} , MSEP estimators.

ˆ) and CV(R̂)
Table 3. Estimators of R̂, MSEP (R
alpha = 0
ALL AV AEHL 5 Years AV 3 Years AV Median
Item/method (1) (2) (3) (4) (5)
R̂ 93 643 65 868 75 886 68 645 54 059
MSEP(R̂ ) 1/2
92 549 88 105 101 643 113 904 105 786
CV(R̂) 99% 134% 134% 166% 196%
(1)/(1) (2)/(1) (3)/(1) (4)/(1) (5)/(1)
R̂ (%) 100% 70% 81% 73% 58%
MSEP(R̂) 1/2
(%) 100% 95% 110% 123% 114%

236 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 236 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

6.2. Robust estimation in GMCL model and Ṽar( f˜k) respectively. These two quantities can
be derived by numerous techniques described in
From the previous two numerical examples (MCL
the literature, such as: M-estimation, Lp estimation,
vs. GMCL results), we observe that, in general, the first
etc. (Huber and Ronchetti 2009) or trimmed mean
approach underestimates and second overestimates
(Jeng 2010).
the MSEP of claims reserves (see Tables 2 and 3).
Step 2. For every k = 1, . . . , I – 1, we find αk
In this section we present an intermediate solution
by solving the following equation fˆk = f˜k, where fˆk is
for our general problem that allows us to evaluate the
given in equation (4.2), namely
estimation of MSEP of reserves in case of develop-
ment factor selection. This go-between solution is I −k
based on the robust statistics in estimation of model ∑ Ciα,k Fi,k
k

parameters fk and sk. The term robust statistics is meant i =1

I −k = fk . (6.1)
in the sense of Huber and Ronchetti (2009). ∑C αk
i, k
i =1
As already mentioned, the assumption GMCL.2
about the conditional variance of Fi,j allows us to
The procedure to select the consistent αk together
estimate the factors fk in the framework of linear
with the problem of existence of solution of equation
regression obtained by the means of weighted least
(6.1) is treated in Murphy, Bardis, and Majidi (2012)
squares procedure (see Murphy, Bardis, and Majidi
(see Lemma 1 and the comments that follow it).
2012 and the references therein). Although these
Step 3. For every k = 1, . . . , I – 2, we find βk by
estimators are easy to compute and have excellent
solving following equation: V̂ar( fˆk) = Ṽar( f˜k), where
theoretical properties (see Proposition 4.1), they rely
V̂ar( fˆk) is given in (A.8), namely,
on quite strict assumptions, and their violation may
lead to useless results. I −k
γ2
One possible solution to overcome this difficulty
I −k
∑ C jβ,k
1
( )  ( fk )
k
2
∑ Ciβ,kk Fi,k − fˆk
j =1
is to use robust estimation techniques. The idea of i
i, k
2 = Var
robust statistics is to account for certain deviations I − k − 1 i =1  I −k 
 ∑ γ j,k 
from idealized model assumptions. Typically, robust j =1

methods reduce the influence of outlying observation (6.2)

on the estimator.
We take the following assumption in our general The parameters fˆk are given in equation (4.2) with
framework of GMCL model: gg,j(t) := t αj and gd,j(t) := αk estimated in Step 2. For k = I – 1, since only one
t βj, with αj,βj ∈  to be estimated. This means that observation is available in our data, βI–1 need to
γi,j := C i,jαj and di,j := C i,jβj. be estimated by other approaches. The limits for the
The following algorithm shows how one can values of Ṽar( f˜k) for which the solution of equation
estimate the parameters αj for j = 1, . . . , I – 1 and (6.2) exists are presented in Appendix E.
βk for k = 1, . . . , I – 2. As can be seen, the presented
method is based on a similar principle to the well 6.2.2. Numerical example
known moment estimation method from point esti- In the present example we consider only the median
mation theory. method already presented in previous numerical appli-
cations. We concentrate on that particular method
6.2.1. Algorithm for fitting because our goal is not to present the extensive case
`k and ak parameters study but rather to illustrate the general principle and
Step 1. We select the robust estimators for fk the main steps of this application. Observe that the
and its variance. We denote these estimators by f˜k method AEHL can be treated by the theory of robust

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 237

15080-07_Sloma-2ndPgs.indd 237 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

Table 4. Estimation of f˜k and s.e(f˜k) using LAD technique corresponding to sample median method
k 1 2 3 4 5 6 7 8 9
f˜k 4,2597 1,5992 1,1635 1,1657 1,1318 1,0335 1,0333 1,0180 1,0092
s.e(f˜k) 1,7974 0,1686 0,1411 0,0270 0,0472 0,0251 0,0065 0,0129 —
CV(f˜k) 42,2% 10,5% 12,1% 2,3% 4,2% 2,4% 0,6% 1,3% —

Table 5. Estimation of parameters `

ˆ j and âj
j 1 2 3 4 5 6 7 8 9
α̂ j 0,5204 1,4073 0 1,5852 –0,3835 0 1,0022 0 0,0000
β̂j 0,6605 0,3120 0 0,7207 1,9501 0 –2,7733 0 —

estimation by means of trimmed mean estimators k = 4, and k = 8 the parameters αk and βk need to be
(Jeng 2010). specified in different way. This could be done using
To apply the above fitting algorithm for sample any other approach that is being judged appropriate
median method, we use the LAD (least absolute by the actuary performing estimation. In our case,
deviation) estimation procedure. The theoretical we put α3 = β3 = α6 = β6 = α8 = β8 = 0 to have from
framework of LAD is presented in Appendix D. one hand the optimal properties (see Proposition 4.1)
The values of f˜k are given by computing the sample but also to be consistent with our choice of α = 0
median from Fi,k as described in Section 3.5. The in our two previous numerical applications (see
standard errors of f˜k are obtained via bootstrap tech- Sections 3.5 and 6.1).
niques. In Table 4 we present the numerical values The estimation of parameters αj and βj are stated
of f˜k, s.e(f˜k) := Ṽar( f˜k)1/2 and CV( f˜k) := s.e( f˜k)/f˜k. Note in Table 5. The values of α̂ j and β̂ j for which we arbi-
that the last value of s.e( f˜k) cannot be estimated from trarily put 0 are indicated with bold font characters.
the data for the reasons discussed in the case of sI–1 The MSEP and claims reserves amount estimators
estimation in Section 3.2 (only one observation avail- are stated in Table 6. Observe that the robust estima-
able). This why we do not fit the parameter βI–1 via tion is a good compromise between the method with
equation (6.2), but we put βI–1 := αI – 1. the same weights (see Section 3.5) and the method
The standard deviations of f˜k from Table 4 were where we use all link ratios in sk estimation (see
obtained by using the rq function integrated in free proxy method in Section 6.1).
R software. Note that the value of f˜k given by this
function are slightly different from those presented 6.3. Validation of results
in Section 3.2. This is probably due to the optimiza- from reserving softwares
tion algorithm that is used in R. Given that these dif- The next interesting and extremely important
ferences are insignificant, we decided to present in application of our GMCL model is the possibility of
Table 4 the same numerical values of the estimators
f˜k as given in Section 3.2. The corresponding R code
is available on request from the author. Table 6. Median method with robust estimation

As mentioned in the algorithm and in Appendix E, Median

the solutions of (6.1) and (6.2) are not always avail- Item/method MCL Robust Proxy

able. For instance, in our example, there is no solu- R̂ 54 059 63 165 54 059

tion of equation (6.1) for k = 3,4 and no solution of MSEP(R̂)1/2 14 786 40 312 105 786
CV(R̂) 27% 64% 196%
equation (6.2) for k = 8. This means that for k = 3,

238 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 238 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

validating the results from industry reserving soft- results we used 100 000 simulations. We begin our
ware. The stochastic chain-ladder type methods are analysis with the classic chain-ladder method in
used to evaluate the economic risk capital required which the estimators of fk are the all volume weighted
by Solvency II for so-called reserve risk. In fact, this average and are consistent with Mack (1993). More
capital requirement for reserve risk is computed as precisely, with the hypothesis of the MCL method
the 99.5th percentile (value at risk) of run-off result with α = 1, we compare the estimate of MSEP
distribution (profit/loss on reserves over one year). obtained by these two techniques: bootstrap from
This means that Solvency II defines the reserve risk ResQ and explicit formula given in MCL approach.
in one-year time horizon, which is different from the The corresponding numerical values are respectively:
standard approach considering the distribution of the 27 150 (see ResQ(Boot) (3) in Table 7) and 26 909
ultimate cost of claims. (see ResQ(MCL) (4) in Table 7). We observe a good
However, one of the methods to derive the one-year convergence for bootstrap (the relative error is less
reserve risk is based on simple scaling of ultimate than 1%). We consider now the different estimator
view. This technique is based on using the results of fk computed as a simple arithmetic average of
of Merz and Wüthrich (2008), which is currently individual link ratios Fi,j. This is equivalent to taking
a popular methodology throughout the market and α = 0 in the MCL framework. In that case, we
taken from the latest technical literature on this topic. observe that the estimates of MSEP for both methods
The empirical loss distribution in ultimate view is become divergent: 75 656 (see ResQ(Mack) (2) in
often derived by using the bootstrap techniques and Table 7) and 58 475 (see ResQ(Boot) (1) in Table 7).
Monte Carlo simulations. The first technique is used This is due to the fact that the ResQ(Mack) method
to evaluate the estimation error and the second to is obtained by approximation based on the MCL for-
approximate the process variance. This kind of boot- mula with α = 1. In fact, according to the technical
strap approach is also available in ResQ software, documentation, the ResQ estimates of parameters
which is used worldwide within the P&C insurance fk and sk in the bootstrap approach are of the form
market. The question is how to validate the results (up to multiplicative constant for bias reduction):
from bootstrap method provided by reserving tools 1 1
fˆk = ∑I–k
i=1Fi,k and s k =
2
∑I–k ˆ 2
i=1Ci,k(Fi,k – fk) .
such as ResQ. One of the possible solutions is to I−k I − k −1
compare the estimation of the first two moments It is easily seen that these estimators are consistent
of loss distribution from bootstrapping (based on with our general approach with α = 0 and β = 1 (see
simulations) with the estimators of reserves and (4.2) and (4.4) in Section 4). The MSEP estimator
MSEP of reserves obtained by the explicit formulas. is equal to 59 065 (see GMCL (5) in Table 7). This
For the sake of simplicity, we assume that there is shows that GMCL method allows one to validate
no factors selection (all weights wi,j are fixed to 1). the results and detect the incoherences. Effectively,
We use the RAA run-off triangle and we present the the choice of estimators in ResQ for the case α = 0
numerical results in Table 7. For all bootstrapping is not optimal in sense of Proposition 4.1. It remains

Table 7. Comparison of ResQ estimators of R̂ and MSEP(R̂) with MCL and GMCL models
alpha = 0,
alpha = 0 alpha = 1 beta = 1
ResQ(Boot) ResQ(Mack) ResQ(Boot) MCL GMCL
Item/method (1) (2) (3) (4) (5)
R̂ 93 630 93 643 52 204 52 135 93 643
MSEP(R̂)1/2 58 475 75 656 27 150 26 909 59 065

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 239

15080-07_Sloma-2ndPgs.indd 239 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

unknown whether this is deliberate or whether this is Appendix: Mathematical Proofs

just a proxy approach that was judged correct.
In regards to the approximation ResQ(Mack) (2), We present here the proofs of our main results.
this shows that in construction of the proxy methods Most of them are derived by simple rewriting the
we cannot just take the MSEP formula for α = 1 as techniques applied in Mack (1993), Mack (1994).
a starting point. Indeed, the MSEP formula changes
if we modify the estimates of fk because the variance A.1. Proof of Result 5.1
of fk is not the same, so it is not enough to plug in Due to the general rule E(X – c)2 = Var(X) + (EX – c)2
the new estimators of Ci,I, fk and sk in the MSEP for- for any scalar c we have
mula (5.4) with α = β = 1. This lack of understanding
of this principle could be a reason of taking the
( )
2
msepCˆiI DI (CiI ) = E  Cˆ iI − CiI DI 
no optimal hypothesis in bootstrap ResQ(Boot) (1)  
method.
( ).
2
= Var (CiI DI ) + E (CiI DI ) − Cˆ iI
Finally, we observe that the results of ResQ(Boot)
(1) method validate our explicit formula for estimation (A.1)
of MSEP of claims reserves in the framework of our
To estimate Var(Ci,I DI) we use the following
GMCL model.
Lemma 9.1. For i = 2, . . . , I, we have,
7. Conclusion
I −1
 Ci2,l  I −1

In this paper we presented a general flexible tool Var (Ci,I DI ) = ∑ E DI  σ l2 ∏ f k2 . (A.2)

l = I +1− i  δ i,l  k =l +1
for stochastic loss reserving and its variability. We
developed our GMCL model to quantify the variabil-
The proof of Lemma A.1 is postponed to
ity of reserves in the context of selecting development
Appendix A.5.
factors in the framework of the stochastic chain-ladder
 Ci2,l 
method. Note that the estimation of E  DI  from equa-
 δ i,l 
We provided the theoretical and flexible back-
tion (A.2) is a crucial part of this proof. We choose
ground which covers some practices of actuaries and
Cˆ i2,l
industrial providers of reserving softwares. to estimate this term by . This is due to the obvious
δˆ i,l
Finally, we showed the way of bridging the chain 2
C
ladder model and the robust estimation techniques. observation that i,l
is an unbiased estimate of
Our results can be applied in other approaches based δ i,l
on chain-ladder framework like: multivariate chain-  Ci2,l 
E   and from the basic property of conditional
ladder, univariate and multivariate Bayesian chain-  δ i,l 
ladder, etc. One can derive the similar results in  C2  C2 
expectation, namely: E  E  i,l DI   = E  i,l  .
the context of one-year reserve risk for Solvency II   δ i,l   δ i,l 
purposes. This topic will be treated in our forth- It is worth noting here that, in the case where
coming paper. Some partial results can be found in di,l := C αi,l with α ∈ , in Saito (2009), the author
Sloma (2014) and Sloma (2011). used the same technique of estimation without giving
any justification or reason for that (see proof of
Lemma 4 and Estimate 8). Similarly, in the case
Acknowledgments
where di,l := wi,l • C αi,l with α ∈ {0, 1, 2} and wi,l ∈ [0, 1],
The author want to thank the reviewers for their we find the same estimator in Mack (1999). More
helpful comments and suggestions. precisely, the author claims (without proving) that

240 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 240 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

∑ I–1
l=I+1–iVar(Ci,l+1DI)∏ k=l+1 f k can be estimated via the
( E (C )
I–1 2
DI ) − Cˆ i,I
2

(s.e.(Fi,l))2/fˆ2l , where (s.e.(Fi,l))2 is an

i, I
quantity Ĉi,I∑ l=I+1–i
I–1

( )
2
estimate of Var(Fi,lCi,1,...,Ci,l). Indeed, this is achieved = Ci2,I +1−i fI +1−i i . . . i fI −1 − fˆI +1−i i . . . i fˆI −1 . (A.4)
 Ci2,l  Cˆ i2,l
if we estimate E  DI  by .
α
 wi,l i Ci,l  wi,l i Cˆ iα,l As can be easily seen, this expression cannot be
However, in Murphy, Bardis, and Majidi (2012), estimated by replacing fk with fˆk. In order to esti-
the authors used different approach based on normal mate the right hand side of (A.4) we use the same
approximation. approach as in Mack (1993), Mack (1994). Saito
Note that in the Section 6.3 we obtained that the (2009) followed the same technique of estimation.
above estimator of Var(Ci,l+1DI) is consistent with that However, in Murphy, Bardis, and Majidi (2012) we
provided by bootstrap technique from ResQ software. can find different approach which was also presented
This is shown for the particular run-off triangle in Buchwalder et al. (2006a). It is worth noting that
and the assumption that Ci,j are gamma-distributed in the paper of Mack, Quarg, and Braun (2006) the
random variables. It would be interesting to perform authors criticised the approaches of Buchwalder
the extensive simulation study in order to examine et al. (2006a) and showed that the estimate of estima-
the exactitude of this estimate with other data and tion error form Mack (1993) is hard to be improved
probability distributions. (see also Buchwalder et al. 2006b). As the answer for
 Ci2,l  Cˆ i2,l the criticism of Mack on article of Buchwalder et al.
We apply now Lemma A.1 with Eˆ  DI  =
 δ i,l  δˆ i,l (2006a), the authors provided the bounds for estima-
and by replacing the unknown parameters fk et s 2k tion error and claimed that the Mack estimator, in
with their estimators fˆk and ŝ k2. Together with the some particular cases, is closed to these bounds (see
equality Ĉi,l = CI+1–i∏ l–1 ˆ Wüthrich, Merz, and Bühlmann 2008). This should
k= I+1–i f k (see Proposition 4.1 (v))
we conclude be confirmed by performing the extensive simulation
study to quantify the different approaches of error
I −1
 Ci2,l  I −1 estimation in stochastic Chain-Ladder framework.
Var (Ci,I DI ) = ∑  δ DI  σ l2 ∏ f k2
E
l = I +1− i  i,l  k =l +1 We define,

I −1
Cˆ i2,l 2 I −1 ˆ 2 F = fI +1−i i . . . i fI −1 − fˆI +1−i i . . . i fˆI −1
= ∑ ˆ σˆ l ∏ f k
l = I +1− i δ i,l k = l +1
= SI +1−i + . . . + SI −1 , (A.5)
I −1
σˆ 2 l −1 I −1
= ∑ l Ci2,I +1−i ∏ fˆ k2 i ∏ fˆ k2
ˆ
l = I +1− i δ k = I +1− i k = l +1 with
i,l

I −1
σˆ l2 fˆ l2 I −1 ˆ 2
= Ci2,I +1−i ∑ ∏ fk Sk = fˆI +1−i i . . . i fˆk −1 fk fk +1 i . . . i fI −1
l = I +1− i δˆ k = I +1− i
i,l

− fˆI +1−i i . . . i fˆk −1 fˆk fk +1 i . . . i fI −1

I −1 I −1
σˆ l2 fˆ l2
= Ci2,I +1−i i ∏ fˆ k2 ∑
k = I +1− i l = I +1− i δˆi,l ( )
= fˆI +1−i i . . . i fˆk −1 fk − fˆk fk +1 i . . . i fI −1 . (A.6)
I −1
σˆ l2 fˆ l2
= Cˆ i2,I ∑ ˆ . (A.3) This yields
l = I +1− i δ i,l

F 2 = ( SI +1−i + . . . + SI −1 )2
We now turn to the second summand of the
I −1 I −1 I −1
expression (A.1). Because of Proposition 4.1 (iv) = ∑ Sk2 + 2 ∑ ∑ S j Sk . (A.7)
and (v) we have, k = I +1− i k = I +1− i j < k

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 241

15080-07_Sloma-2ndPgs.indd 241 4/25/19 11:27 AM

 Variance Advancing the Science of Risk

We estimate F 2 using the following A.2. Proof of result 5.2

(overall standard error)
Proposition A.1 (Estimate of F2) Let define, for Following the definition in (2.4), we have
1 ≤ k ≤ I, the set of observed Ci,j up to development
year k, namely  C 
I
msep I ∑ i,I
∑ Ci ,I DI  i =1 
ˆ
i =1

Bk = {Ci, j : i + j ≤ I + 1, k ≤ j} ⊂ DI .
2
 I I

= E  ∑ Cˆ iI − ∑ CiI  DI 
 i =1 i =1  
Then, we can estimate F 2 by
I I I 2
= Var  ∑ Ci,I DI  +  E  ∑ Ci,I DI  − ∑ Cˆ i,I 
I −1 I −1 Var ( fˆk Bk )  i =1    i =1  i =1 
2 = fˆ l2
F ∏ ∑ .
l = I +1− i k = I +1− i fˆ k2
The independence of accident years yields

The proof of Proposition A.1 is postponed to II

Var  ∑ Ci,I DI  = ∑ Var (Ci,I DI ).
Appendix A.5.  i =1  i =1
It remains to determine the estimate of Var( fˆk Bk).
We use the following where each term of the sum has already been calcu-
lated in the proof of the Result 5.1.
Proposition A.2 We assume (4.3). We have Furthermore

I 2
 E  C D  − I Cˆ 
I −k
γ 2   ∑ i,I I  ∑ i,I 
∑δ j, k
i 1{δ ≠ 0}
j, k
i =1 i =1

Var ( fˆk Bk ) = σ 2 j =1 j, k
. (A.8) I 2
=  ∑ ( E (Ci,I DI ) − Cˆ i,I )
k I −k 2
 γ 
 ∑ j,k   i =1 
j =1
I
= ∑ ( E (Ci,I DI ) − Cˆ i,I )( E (C j,I DI ) − Cˆ j,I )
The proof of Proposition A.2 is postponed to i, j

Appendix A.5.
Finally, using (A.4) and Proposition A.2 we estimate Taking together
E(Ci,I DI) – Ĉi,I)2 by I I
 I ˆ D  ∑ Ci ,I  = ∑ msep
msep  ˆ D (CiI )
∑
i =1
CiI I
 i =1  i=2 CiI I
I −k
γ2
∑ δ j,k i 1{δ ≠ 0} I
I −1
σˆ 2 j, k

C 2
i,I +1− i fˆ I2+1−i i . . . i fˆ I2−1 ∑ ˆ 2k
j =1 j, k
I −k 2
+ ∑ 2 i Ci,I +1−iC j,I +1− j Fi Fj ,
k = I +1− i f k  γ  2≤i ≤ j ≤ I
 ∑ j,k 
j =1
with
I −k
γ2
∑ δ j,k i 1{δ ≠ 0}
Fi = fI +1−i i . . . i fI −1 − fˆI +1−i i . . . i fˆI −1 =
I −1
I −1
σˆ 2 j, k
∑ Ski ,
= Cˆ 2i,I ∑ 2k
j =1 j, k
. k = I +1− i
ˆ
k = I +1− i f k
I −k
 γ 
2

 ∑ j,k 
j =1 where

This completes the proof of Result 5.1.  Ski = fˆI +1−i i . . . i fˆk −1 ( fk − fˆk ) fk +1 i . . . i fI −1 .

242 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 242 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

We can determine the estimator of Fi Fj in the analo- We have, for 1 ≤ k ≤ I – 2,

gous way as for F2.
I −k
( I − k − 1) i σˆ 2k = ∑ δ i,k ( Fi,k − fˆk )
2

Proposition A.3. We have i =1

I −k I −k I −k
I −1Var ( fˆk Bk ) = ∑ δ i,k F i2,k − 2 ∑ δ i,k Fi,k i fˆk + ∑ δ i,k fˆk2 .

Fi Fj = ∑ ( fˆI +1−i i . . . i fˆI −1 ) i =1 i =1 i =1
k = I +1− i fˆ k2
Since di,k are s(Ci,k) measurable, we have
i ( fˆ I +1− j i . . . i fˆI −1 ).
I −k
E (( I − k − 1) i σˆ 2k Bk ) = ∑ δ i,k E ( F i2,k Bk )
We finally conclude, from Proposition A.3 i =1

I −k
I − 2 ∑ δ i,k E ( Fi,k i fˆk Bk )
∑ 
2 i Ci,I +1−iC j,I +1− j Fi Fj
i =1
2≤i < j ≤ I
I −k
+ ∑ δ i,k E ( fˆ k2 Bk ).
I I I −1
σˆ 2k fˆ k2 I −k γ l2,k i 1{δ j,k ≠ 0}
= ∑ Cˆ i,I ∑ Cˆ j,I ∑ ∑
i =1
2 2
i .
I −k
 γ  δ l,k
∑
i=2 j = i +1 k = I − i +1 l =1

l =1
l, k
 In the following derivation we use s(Ci,k)-
measurability of γi,k and definition of fˆk from (4.2).
Furthermore, the assumption GMCL.3 implies that Fi,k
A.3. Proof of Proposition 3.1
and Fj,k are independent for i ≠ j. From assumption
i. See Theorem 2 p. 215 in Mack (1993). GMCL.1 and GMCL.2 we easily see that E(F 2i,kBk) =
ii. See discussion on p. 112, Corollary on p. 141 σ 2k
+ f k2 . Taking together,
and Appendix B on p. 140 in Mack (1994). δ i,k
iii. See Appendix E on p. 151 in Mack (1994).
1  I −k 
iv. 
See Theorem 1 p. 215 and discussion after (
E Fi,k i fˆk Bk = ) I −k  ∑ γ j,k i E ( Fi,k i Fj,k Bk )
the proof of Theorem 2 on page 216 in Mack ∑ γ l,k j =1

l =1
(1993).
I −k
 γ i E F2 B + γ 
v. see Appendix C p. 142 in Mack (1994):
1  i,k
( i,k k ) ∑ j,k 
= I −k 
j ≠i

A.4. Proof of Proposition 4.1 ∑ γ l,k  i E ( F B ) i E ( F B ) 
i, k k j, k k
l =1

(i), (iv) and (v), see proofs of (i), (iv) and (v) respec-
1   σ 2k 2
I −k
2
tively in Proposition 3.1. = I −k  γ i,k i  δ + f k  + ∑ γ j,k f k 
(ii). The first part of the statement regarding to ∑ γ l,k i, k j ≠i

l =1
the minimal variance of parameters fk can be
easily derived from the proof of (ii) in Propo- 1  σ 2k I −k

= I −k  γ i, k i
δ i,k
+ f 2
k ∑ γ j,k 

sition 3.1. The rest of the proof is easily seen ∑ γ l,k j=i

from the Proposition A.2. l =1

(iii). Without loss of generality and to avoid the γ i,k

complexity of notation we present the proof for δ i,k
= σ 2k I −k + f k2 . (A.9)
Ik = I – k (for each k, all weights di,k are different
∑ γ l,k
from 0). l =1

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 243

15080-07_Sloma-2ndPgs.indd 243 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

From Proposition A.2 A.5. Proofs of auxiliary results

( ) ( ) ( ( )) A.5.1. Proof of Lemma 9.1

2
E fˆ k2 Bk = Var fˆk Bk + E fˆk Bk
Let define, for 1 ≤ i ≤ I and 1 ≤ j ≤ I, the set of
I −k
γ2 observed data Ci,j for accident year i and up to devel-
∑ δ j,k
j =1 j, k opment year j, namely
= σ 2k 2 + f k2 .
 I −k

 ∑ γ j,k  Ai, j = {Ci,k : 1 ≤ k ≤ j} .
j =1

Taking together we have For l = I + 1 – i, . . . , I – 1,

E (( I − k − 1) i σˆ 2k Bk ) Var (Ci,l DI ) = Var (Ci,l +1 Ai ,I +1−i )

I −k I −k E [Var (Ci,l +1 Ai,l −1 ) Ai,I +1−i ]
= ∑ δ i,k E ( F i2,k Bk ) − 2 ∑ δ i,k E ( Fi,k i fˆk Bk )
i =1 i =1
+ Var [ E (Ci,l +1 Ai,I −1 ) Ai,I +1−i ]
I −k
(
+ ∑ δ i,k E fˆ k2 Bk )  Ci,l
2

Ai,I +1−i  + Var [ flCi,l Ai,I +1−i ]
i =1 = E  σ l2
γ i,k  δ i,l 
 
I −k
 σ 2k  I −k  δ   Ci,l
2

= ∑ δ i,k  + f k2  − 2 ∑ δ i,k  σ 2k I − k Ai,I +1−i  + f l2Var [Ci,l Ai,I +1−i ]
i, k
+ f k2  = σ E 2

i =1  δ i,k  i =1
 ∑ γ i,k 
l
 δ i,l 
 i =1 
2
 Ci,l 
 γ I −k  2 = σ l2 E  DI  + f l2Var [Ci,l DI ].
∑δ j, k δ
 i,l 
I −k  
+ ∑ δ i,k  σ 2k
j =1 j, k
2 + f 2
k  (A.11)
i =1   I −k  
  ∑ γ j,k  
j =1 We multiply the both sides by ∏ I–1 k=l+1 f k with the con-
2

I −k I −k vention that an empty product equals 1. Taking the

= ( I − k ) σ 2k + f k2 ∑ δ i,k − 2σ 2k − 2 f k2 ∑ δ i,k
i =1 i =1 sum over l = I + 1 – i, . . . , I – 1, we obtain
I −k I −k
γ 2
I −1 I −1
∑ δ i,k ∑ δ j, k
I −k ∑ Var (Ci,l +1 DI ) ∏ f k2
+ f k2 ∑ δ i,k
i =1 j =1 j, k
+ σ 2k − 2
l = I +1− i k = l +1

 I −k
 i =1
 ∑ γ j,k 
2
I −1
 C i,l  I −1
j =1 = ∑ σ l2 E  δ DI  ∏ f k2
l = I +1− i  i,l  k =l +1
 I − k I − k γ j,k
2

 ∑ δ ∑ 
I −1 I −1
∑ Var [Ci ,l DI ] f l2 ∏ f k2Var (Ci ,I DI )
i, k
i =1 j =1 δ j, k +
= ( I − k − 1) σ 2k + σ 2k  2 − 1 . (A.10) l = I +1− i k =l
  I −k  
  ∑ γ j,k   I −2 I −1
 j =1  + ∑ Var (Ci ,l +1 DI ) ∏ f k2
l = I +1− i k = I +1
Finally 2
I −1
 Ci,l  I −1
= ∑ σ l2 E  DI  ∏ f k2
E ( σˆ 2k − σ 2k ) = E [ E [( σˆ 2k − σ 2k ) Bk ]] l = I +1− i  δ i,l  k =l +1
I −1
 I − k I − k γ 2j,k  + Var [Ci,I +1−i DI ]
 ∑ δ i,k ∑  ∏ f k2
σk j =1 δ j, k
2 k = I +1− i
= E  i =1
− 1 .
I − k − 1   I −k 
2
 I −1 I −1
  ∑ γ j,k   + ∑ Var [Ci,l DI ]∏ f k2 . (A.12)
  j =1   l = I + 2−i k =l

244 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 244 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

Since Var[Ci,I+1-i DI] = 0 and from the fact that A.5.4. Proof of Proposition A.3
We find the estimator Fi Fj in the similar way to
I −2 I −1 I −1 I −1 2
the estimator F (see proof of Proposition A.1).
∑ Var (Ci,l +1 DI ) ∏ f k2 = ∑ Var [Ci ,l DI ] ∏ f k2 ,
l = I +1− i k = l +1 l = I + 2− i k =l

E [( Ski )2 Bk ] = fˆI2+1−i i . . . i fˆk2−iVar ( fˆk Bk ) f k2+1 i . . . i f I2−1 .

we finally get the proof of Lemma A.1
For i < j, we have
A.5.2. Proof of Proposition A.1
I −1
Following Mack (1993), Mack (1994), we replace 
Fi Fj = ∑ fˆI +1− j i . . . i fˆI −i
S k with E(S k2Bk) and Sj Sk, with E(Sj SkBk). This means
2 k = I +1− i

that we approximate S k2 and Sj Sk by varying and aver- i fˆI2+1−i i . . . i fˆk2−iVar ( fˆk Bk ) fˆk2+1 i . . . i fˆI2−1
aging as little data as possible so that as many values
Ci,k from data observed are kept fixed. Due to Propo- I −1 Var ( fˆk Bk ) ˆ
= ∑ fI +1− j i . . . i fˆI −i
sition 4.1 (i) we have E( fˆk – fk) = 0 and therefore k = I +1− i fˆk2
E(Sj SkBk) = 0 for j < k because all fr, r < k, are scalars
under Bk. Since E(( fk – fˆk)2Bk) = Var( fˆkBk) we obtain i fˆ I2+1−i i . . . i fˆk2−1 i fˆk2 i fˆk2+1 i . . . i fˆI2−1
from (A.6)
I −1 Var ( fˆk Bk ) ˆ
= ∑ ( fI +1− j i . . . i fˆI −1 )
E ( S 2k Bk ) = fˆI2+1−i i . . . i fˆk2−i Var ( fˆk Bk ) f k2+1 i . . . i f I2−1 . k = I +1− i fˆk2

Taken together, we have replaced F2 = ∑ I–1

i ( fˆI +1−i i . . . i fˆI −1 ) . (A.13)
k=I+1-i Sk
2

with ∑ I–1
k=I+1-i E(S k Bk) and the unknown parameters
2

are replaced by their estimators. Altogether, we esti-

mate F 2 by

I −1 Var ( fˆk Bk )
∑ fˆI2+1−i i . . . i fˆk2−1 fˆk2 fˆk2+1 i . . . i fˆI2−1 .
k = I +1− i fˆk2

A.5.3. Proof of Proposition A.2

From definition of fˆk in (14), we have

I −k
 γ F 
∑
 j =1 j,k j,k 
Var ( fˆk Bk ) = Var  I − k Bk 
 ∑ γ j,k 
 j =1 
I −k
∑ γ 2j,kVar ( Fi,k Bk ) i 1{δ j,k ≠ 0}
j =1
= I −k 2 ,
 γ 
 ∑ j,k 
j =1

where the second equality is due to the Bk-measurability

of γj,k, the assumption (4.3) and the convention that the
product of 0 and ∞ equals to 0.

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 245

15080-07_Sloma-2ndPgs.indd 245 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

B. Data
We present in Table B.8 the triangle of RAA data analysed in Mack (1994) and Murphy, Bardis, and
Majidi (2012).

Table B.8. RAA run-off triangle (cumulative payments)

Development Year j
Accident
Year i 1 2 3 4 5 6 7 8 9 10
1 5 012 8 269 10 907 11 805 13 539 16 181 18 009 18 608 18 662 18 834
2 106 4 285 5 396 10 666 13 782 15 599 15 496 16 169 16 704
3 3 410 8 992 13 873 16 141 18 735 22 214 22 863 23 466
4 5 655 11 555 15 766 21 266 23 425 26 083 27 067
5 1 092 9 565 15 836 22 169 25 955 26 180
6 1 513 6 445 11 702 12 935 15 852
7 557 4 020 10 946 12 314
8 1 351 6 947 13 112
9 3 133 5 395
10 2 063

C. Individual link ratios of RAA run-off triangle

Table C.9. Individual link ratios Fi,j (age-to-age factors) of run-off triangle RAA
AY 1–2 2–3 3–4 4–5 5–6 6–7 7–8 8–9 9–10
1 1,650 1,319 1,082 1,147 1,195 1,113 1,033 1,003 1,009
2 40,425 1,259 1,977 1,292 1,132 0,993 1,043 1,033
3 2,637 1,543 1,163 1,161 1,186 1,029 1,026
4 2,043 1,364 1,349 1,102 1,113 1,038
5 8,759 1,656 1,400 1,171 1,009
6 4,260 1,816 1,105 1,226
7 7,217 2,723 1,125
8 5,142 1,887
9 1,722

D. LAD estimator
{ } {∑ }
n n m
min ∑ ei = min yi − ∑ xijβ j ,
β β
The least absolute deviation (LAD) method or L1 i =1 i =1 j

(also known as Least Absolute Value (LAV)) method

is a widely known alternative to the classical least where ei := yi – ∑ mj xij βj, i = 1, 2, . . . , n and j = 1,
squares (LS) or L2 method for statistical analysis of 2, . . . , m. Unlike the LS method, the LAD method
linear regression models. Instead of minimizing the is not sensitive to outliers and produces robust esti-
sum of squared errors, it minimizes the sum of abso- mates. LAD method is reduced to a linear program-
lute values of errors. More precisely, in the context ming problem and the computational difficulty is now
of linear regression model, estimates are found by entirely overcome by the availability of computing
solving the following optimisation problem power and the effectiveness of linear programming.

246 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 246 4/25/19 10:47 AM

 Generalized Mack Chain-Ladder Model of Reserving with Robust Estimation

Least absolute values (LAV) regression is very resis- N N

h′ (β ) := ∑ ∑ bi c j i ln ( di, j ) i d βi, j
tant to observations with unusual values in data. i =1 j = i +1
In the numerical example presented in Section 6.2.2,
N N
we used one-dimensional (m = 1) LAD procedure +∑ ∑ b j ci i ln ( d j,i ) i d βj,i ,
where we took, for each column k of run-of triangle, i =1 j = i +1

yi := Fk,i and xi1: = 1 for all i.

One more thing merits mentioning here. In the Since ln(di,j) > 0 and ln(dj,i) < 0, the first derivative h′
simple one-dimensional case (m = 1) the LAD esti- has a limit in –∞ and ∞ if β tends to –∞ and ∞
mator yields to the sample median (see Abur and respectively. In addition, the second derivative h″ is
Expósito 2004, p. 141). given by

E. Robust estimation—Limits in estimation N N

h′′ (β ) := ∑ ∑ bi c j i ( ln ( di, j )) d βi, j
2
i
of a parameters i =1 j = i +1

We want to examine the existence of solution of N N

+∑ ∑ b j ci i ( ln ( d j,i )) d βj,i .
2
equation (6.2). In this purpose, we study the proper- i
i =1 j = i +1
ties of the flowing type of functions,

N N Given that d βi,j and d βj,i are strictly positive functions

hk (β k ) :=  ∑ ai−β  k β 
k

, k bi, k   ∑ ai, k ci, k  , and all coefficients are positive, the second derivative
 i =1   i =1 
of h is strictly positive. This means that first deriva-
with ai,k := Ci,k, bi,k := γ 2i,k/(∑ i=1
I–k
γi,k)2, ci,k := 1/(I – k – 1) tive of h is increasing function. Together with the pre-
( fˆk – Fi,k) and Nk := I – k. In the sequel, without the vious facts it implies that h has an absolute minimum.
loss of generality, we omit the index k corresponding In consequence, the equation (23) has zero, one or
to the column of run-off triangle. Thus, we consider two solutions.
the function In the case where two solutions of opposite sign
exist, the actuary should decide which one cor­
N N
h (β ) :=  ∑ ai−β bi   ∑ aiβci  , responds better to the considered line of business.
 i =1   i =1  In fact, as mentioned in Murphy, Bardis, and Majidi
(2012), the choice of negative solution does not
with ai ≥ 0, 0 ≤ bi ≤ 1 and ci ≥ 0. We rewrite the
seem to be unreasonable in some situations. This
function h as follows:
issue is out of scope of this paper. In our numerical
N N example the solution is determined by Excel tool
h (β ) := ∑ ∑ ( ai−β bi ) ( aβj c j ) . called solver.
i =1 j =1

We easily observe that the function h tends to ∞ as β References

tends to ∞ or –∞. Let us define, di,j: = (ajk/aik), for i < j,
Abur, A., and A. Expósito, Power System State Estimation:
and where the indices ik < jk are such that (ajk/aik) > 1. Theory and Implementation, Boca Raton, FL: CRC Press,
Then, by simple decomposition of double sum, we get: 2004.
Barnett, G., and B. Zehnwirth, “Best Estimates for Reserves,”
N N N N N Proceedings of the Casualty Actuarial Society 87, Part 2,
h (β ) := ∑ bi ci + ∑∑ bi c j i d βi, j + ∑ ∑ b j ci i d βj,i , 2000, pp. 245–321.
i =1 i =1 j = i +1 i =1 j = i +1
Blumsohn, G., and M. Laufer, “Unstable Loss Development
Factors,” Casualty Actuarial Society E-Forum, 2009, pp. 1–38.
where by our notation di,j > 1 and dj,i < 1. By the Buchwalder, M., H. Bühlmann, M. Merz, and M. V. Wüthrich,
straightforward computations it is easy to show that “The Mean Square Error of Prediction in the Chain Ladder

VOLUME 12/ISSUE 2 CASUALTY ACTUARIAL SOCIETY 247

15080-07_Sloma-2ndPgs.indd 247 4/25/19 10:47 AM

 Variance Advancing the Science of Risk

Reserving Method (Mack and Murphy Revisited),” ASTIN Merz, M., and M. V. Wüthrich, “Modelling the Claims Develop-
Bulletin 36: 2, 2006a, pp. 521–542. ment Result for Solvency Purposes,” Conference Paper, ASTIN
Buchwalder, M., H. Bühlmann, M. Merz, and M. V. Wüthrich, Colloquium, Manchester, July 2008.
“The Mean Square Error of Prediction in the Chain Ladder Murphy, D. M., “Unbiased Loss Development Factors,” Insurance
Reserving Method—Final Remark,” ASTIN Bulletin 36, 2006b, Mathematics and Economics 18: 3, 1996, p. 228.
pp. 553–553. Murphy, D., M. Bardis, and A. Majidi, “A Family of Chain-Ladder
Huber, P. J., and E. M. Ronchetti, Robust Statistics, 2009. Factor Models for Selected Link Ratios,” Variance 6, 2012,
Jeng, H.-W., “On Small Samples and the Use of Robust Estimators pp.143–160.
in Loss Reserving,” Casualty Actuarial Society E-Forum, Saito, S., “Generalisation of Mack’s Formula for Claims Reserv-
Fall 2010. ing with Arbitrary Exponents for the Variance Assumption,”
Mack, T., “Distribution-Free Calculation of the Standard Error of Journal of Mathematics for Industry 1, 2009, pp. 7–15.
Chain Ladder Reserve Estimates,” AST1N Bulletin 23, 1993, Sloma, P. Contribution to the Weak Convergence of Empirical
pp. 213–222. Copula Process. Contribution to the Stochastic Claims Reserv-
Mack, T., “Measuring the Variability of Chain Ladder Reserve ing in General Insurance. PhD thesis, Université Pierre et Marie
Estimates,” Casualty Actuarial Society Forum 1, 1994, Currie, 2014.
pp. 101–182. Sloma, P., “General Model for Measuring the Uncertainty of the
Mack, T., “The Standard Error of Chain Ladder Reserve Estimates Claims Development Result (CDR),” Proceedings of the Actu-
Recursive Calculation and Inclusion of Tail Factor,” ASTIN arial and Financial Mathematics Conference, February 10–11,
Bulletin 29: 2, 1999, pp. 361–366. 2011.
Mack, T., G. Quarg, and C. Braun, “The Mean Square Error of Pre- Wüthrich, M. V., M. Merz, and H. Bühlmann, “Bounds on the
diction in the Chain Ladder Reserving Method: A Comment,” Estimation Error in the Chain Ladder Method,” Scandinavian
ASTIN Bulletin 36: 2, 2006, pp. 543–552. Actuarial Journal 2008: 4, 2008, pp. 283–300.

248 CASUALTY ACTUARIAL SOCIETY VOLUME 12/ISSUE 2

15080-07_Sloma-2ndPgs.indd 248 4/25/19 10:47 AM